Hi.
I am looking for a training model that is best fit for me. My system only has like 6GB vram max, and it's only safe for the training to take up at around 5.7 GB. I tried several models, like original and dfl-sae, but I would want to be sure before waiting for weeks of training.
I am looking for a model that can get very good results, but is possible to run on a 6GB vram. It is ok to have a slow training rate, like 1 batch per iteration to work, but to me having good results is king, time cost of training doesn't matter.
Then it's the swap issue. My previous swaps looked like someone wearing a mask: The difference in skin color is too obvious, and it's too blur. For the context, the model was a dfl-sae with 16 batches per iterations at 75000 iterations, and 128 px I guess? I know this ain't enough, so I think it's more of a training or model inefficiency rather than swap, so if the first question could solve this just forget it.
Thanks.